AGENTS.md instructions for /home/maxwell-rosen/postgkyl

0. Locality of reasoning. Every principle below is a projection of one axiom: a reader must be able to understand a fragment without the whole program. Whatever keeps a local conclusion sound — a frozen record, an honest signature, a stated law — is doctrine. Whatever forces a global search — ambient state, a leaky layer, a second copy of a fact — is the enemy.

Data — what it does, and what it may say

I. Data is inert. Functions transform. No objects that know things and do things. Data is a frozen record. Behavior is a function that takes data in and returns data out. If you're reaching for inheritance, you've taken a wrong turn.

II. Make illegal states unrepresentable. The shape of a datum is its strongest invariant. Constructors refuse invalid states; a checked fact becomes a type; downstream never re-proves what upstream established. Parse, don't validate.

Functions — one idea, honestly declared

III. A function is one idea. It takes exactly what it needs and returns exactly what it computes. If the signature has two concepts in it, you have two functions.

IV. The signature tells the whole truth. Inward: if something needs a value, it receives it as a parameter — no spooky action at a distance, no stringly-typed interfaces, no implicit state. Outward: same inputs, same outputs; effects and failure appear in the type, not in the fine print. Pure core, effects at the edges.

Knowledge — one home per fact

V. Every fact has one home. One authoritative representation of each decision and each piece of knowledge; everything else inherits or is derived mechanically — never maintained by hand in parallel. Configuration is decided once, at the highest level, and threaded down; no module ever decides its own context. If the design and the implementation can disagree, you have two sources of truth and zero.

Layers — what above, how below

VI. Separate what from how. Logic and machinery are different concerns with a hard boundary. The layer that says what to compute should be readable by someone who has never seen the machinery underneath. The layer that says how lives below, stays below, and nothing leaks up from it.

VII. Notation is execution; lowering is transliteration. Looking up: the spec layer reads like the math or logic it implements — when notation is the executable object, not a comment beside it, bugs have nowhere to hide. Looking down: the layer that executes the spec reproduces it exactly — nothing added, nothing dropped, nothing reinterpreted; no opinions, no defaults, no helpful conversions. If the lowering changes anything, the spec is a lie.

Abstraction — earned, and binding

VIII. Earn your abstractions. No abstraction before the second use. Three similar lines is better than a premature helper. The right amount of complexity is the minimum the current task demands — not the current task plus three hypothetical future ones.

IX. An abstraction is a contract. It is defined by what it guarantees, not what it hides. If you can't state what is always true of it — properties a client may rely on without reading the implementation — it isn't an abstraction, it's indirection. Two implementations that honor the contract must be interchangeable; and its outputs stay in its vocabulary, so uses compose.

Verification — formal first

X. Trust the most formal thing first. Types over tests, tests over docs, docs over comments. Invest in whichever layer catches the bug earliest with the least ongoing maintenance cost.

Commands

Boolean API/CLI contract

Every public boolean parameter defaults to False. If the default behavior is enabled, name the parameter for disabling or selecting its inverse (for example, no_show=False, volume=False, or nodal=False) and implement the default behavior under the false branch. The generated CLI mirrors the API parameter name and default exactly; a bare boolean option means True, while an explicit True or False remains accepted. The CLI compiler rejects a public boolean whose API default is not False.

Architecture — a strict, one-way layered DAG

Every folder has one job, and imports point in one direction only (leaves at the bottom). There is no import cycle — this is enforced by a test (see "Import contract"). Arrow = "may import":

Keeping the picture honest: the two diagrams below and the prose after them are a mirror of tests/test_postgkyl.py::_ALLOWED — that dict (and the AST walk that checks every real import against it) is the enforced source of truth; this file is only a readable projection of it. The two can drift (they already had: operations/average.py, operations/eval_at_coord_proj.py, operations/local_poly.py, gdatastate/guards.py, gdata/gdatagroup.py, and gdata/verbs.py existed in the tree before they were added here). Whenever you add a new top-level module file or a new allowed import edge, update _ALLOWED and this section in the same commit — don't let the picture outlive the code it describes.

The two-domain lifecycle (REFACTOR_GKEYLL_FFI.md)

Every dataset lives in one of two backends, discriminated by GDataState.backend:

• "gkyl" (modal domain) — DG coefficients as a native gkyl_array (gpython.GkylArray). Loading lands here. All math runs inside Gkeyll: *// → weak kernels (gkyl_dg_mul_op/div_op), +/- → coefficient lin-combs (gkyl_array_accumulate), scalars → scale/mean-shift, integer powers → repeated weak multiply, full .integrate() → gkyl_array_integrate, partial .integrate(axis=…) → exact gkyl_array_average followed by physical-volume scaling. values is a read-only view; np.asarray/ufuncs/select refuse with ".interpolate() first".
• "numpy" (field domain) — post-interpolate() values as a plain ndarray; the unchanged NumPy stack (select, plot, ufuncs, arithmetic).

interpolate() is the one-way bridge: matrix from Gkeyll's basis functions, applied per cell with NumPy tensordot, returning a new, by-value array.

Every dataset has one value_form (ctx["value_form"]): modal coefficients, nodal values at the basis node_list points, or quad values at Gauss–Legendre points. This is a single, backend-agnostic fact — there is no separate is_modal flag duplicating it; a numpy-backed dataset that was never native (e.g. a plain nodal-basis file read without the Gkeyll library) carries the same three-valued value_form as a gkyl-native one, and every consumer (_require_operable, interpolate, average, …) reads that one key. The capability boundary is modal vs point-values, not gkyl vs NumPy:

• modal — only Gkeyll's DG operations: weak * /, coefficient + -/scalar kernels, .integrate(), .interpolate(). Ufuncs/np.asarray/plot refuse.
• nodal / quad — the values ARE the field at points, so every pointwise NumPy operation is exact and allowed (ufuncs, * / **, np.asarray) — computed on the views, wrapped back native, staying in-value_form — and they plot() directly at their true point locations (non-tensor node sets, e.g. serendipity p2 in 2-D, plot via .to_quad()).

Conversions are never implicit — only .to_modal()/.to_nodal()/.to_quad() change value_form (nodal↔modal exact; quad round-trip exact for degree ≤ 2·num_quad−1); .apply(fn, num_quad=…) is the one-shot modal → quad → fn → project-back spelling (≡ fn(d.to_quad()).to_modal()). Datasets combine only within one value_form. See REFACTOR_GKEYLL_FFI.md §3b.

basis_type, poly_order, and value_form are properties of the data itself — read from a file's header metadata, or set once via pg.load(..., basis_type=..., poly_order=..., value_form=...) / the CLI's bare-filename load (-b/-p/-v) — never re-specified by a downstream verb. .interpolate(), .local_poly(), .average(), .eval_at_coord_proj(), .integrate(), … all read ctx["basis_type"]/ctx["poly_order"]/ctx["value_form"] off the dataset and raise a clear error if a required one is missing; none of them take a basis/poly_order override argument. Loading a dataset that has a spatial grid (ctx["cells"]) without basis_type/poly_order/value_form resolvable (neither in the file header nor given explicitly) warns and defaults to basis_type="serendipity", poly_order=0, value_form="nodal" — the trivial one-point-per-cell basis, since there is no modal structure to assume otherwise. When value_form is defaulted this way, the grid is re-expressed as cell centers (p0 nodal's one point per cell) instead of the reader's raw cell-edge grid, so it lines up one-to-one with values. A dynvector/diagnostic file has no spatial grid and thus no DG basis to speak of, so it is exempt from this defaulting. This is distinct from the readers' own narrower default (gkyl_c_reader.py/gkyl_reader.py: basis_type resolved but no value_form tag in the file -> assume "modal", silently, no warning) — there the data is known to be real DG output (the header has a basis), so the stored numbers are almost certainly modal coefficients already, not point values; that assumption is safe enough not to warn about. The gdatastate.py-level default above only fires when nothing about a basis was found at all.

gdatastate/ — the container (gdatastate/state.py)

GDataState holds one dataset: a nodal grid (list of 1-D edge arrays), values in one of the two backends (gpython.GkylArray or np.ndarray), and metadata in ctx. It is verb-less and imports only downward (io to construct itself, gpython for the backend type, and dg for shared point-value materialization). It owns:

• shape properties (num_dims/num_comps/num_cells/bounds), grid/values,
• backend ("gkyl"/"numpy") and native (the raw GkylArray for the kernels),
• push, clone (backend-aware deep copy via type(self)), and _result(...) — the one "mutate-self vs. emit-new" decision point every verb funnels through,
• pure state readers only: __array__ (refuses on gkyl-backed data), __repr__/__str__, info, is_interpolated. gdatastate/collection.py has flatten_datasets (shared by the multi-dataset entry points). gdatastate/guards.py centralizes the field-domain check (backend == "gkyl" → raise with the standard ".interpolate() first" message) that several operations/diagnostics verbs need but that isn't itself a verb, so it lives here rather than in operations. gdatastate/materialize.py owns the shared conversion from native nodal/quad values to a NumPy-backed state used by every terminal consumer.

api/ — the fluent surface (api/gdata.py, api/load.py)

class GData(GDataState) adds the fluent verb methods (.interpolate(), .select(), .plot(), .save(), .info inherited) and the computing operators (+ - * / **, reflected, __neg__/__abs__, __array_ufunc__). Because it lives above operations, exact verbs are static class-body aliases to their canonical operation — no wrapper, runtime setattr, or lazy import. pg.load(...) returns a GData.

gdata/gdatagroup.py mirrors the same move one level up: class GDataGroup(gdatastate.GDataStateGroup) adds broadcasting — any attribute not defined on the class is resolved by __getattr__, looked up on every member, so a verb call broadcasts across the whole group without a single verb body being duplicated. gdata/verbs.py holds the handful of verbs that combine several datasets and so have no single self to hang off of a class — collect, evaluate, relchange, animate — each a direct alias to the matching operations function; GData and GDataGroup both call through these same module-level functions for their own methods, so the functional and fluent spellings of a multi-dataset verb can never drift apart.

The trick that removes the cycle: operations verbs are typed on GDataState but return the caller's concrete class, because _result builds type(self)(). So operations never needs to import gdata, yet the whole fluent chain stays GData. See HIERARCHY_2.md.

operations/ — the data-transformation library (the single seam)

Flat core verbs have one module each and domain transformations live in named subpackages exposed from operations/__init__.py. Contract: op(data: GDataState, *, ..., inplace=False, tag=None, label=None) -> GDataState. Flat modules hold equation-blind core verbs. Domain subpackages may know the geometry or representation of one equation system while still only transforming/re-expressing data; operations/gyrokinetics/ owns the R-Z and flux-surface projections. Code that interprets components to derive a new physical conclusion belongs in diagnostics/. Implemented: interpolate (the bridge verb: gkyl-backed in, numpy-backed out), select (field-domain only), info, integrate (full or partial; modal data stay inside Gkeyll, with full integration terminal and partial integration returning an exact lower-dimensional modal dataset), represent/apply (the explicit value_form verbs behind .to_modal()/.to_nodal()/.to_quad()/.apply()), arithmetic (binary + apply_ufunc), which dispatches on backend: modal operands → dg.modal kernel calls; numpy operands → the NumPy path; mixed domains or mixed value_forms → error, plus the field-domain analysis verbs (fft, magsq, relchange, mask, collect, grid, val2coord, extract_input, fit (its window=True mode covers growth-rate-style leading-window fits), differentiate, evaluate, map); and the modal-native verbs average (weighted average over a dimension subset via Gkeyll's gkyl_array_average) and eval_at_coord_proj (eval at physical coordinates, projected onto the lower-dimensional basis for the surviving directions) — both terminal-adjacent like represent: they emit a new, lower-dimensional dataset that stays modal/gkyl-native, so it composes with further .to_nodal()/.interpolate()/.average()/.eval_at_coord_proj() calls rather than dropping to NumPy. local_poly bridges modal coefficients to a discontinuity-preserving plotting mesh. operations.plot/animate/plotly/plotly_animate/pyvista are direct aliases of their canonical render callables. All terminal consumers share gdatastate.materialize_point_values, so the point-value capability rule has one home. Verbs wrap the layers below; they don't reimplement.

diagnostics/ — equation-specific physics (COMPOSITION, above gdata)

The layer that knows what the numbers mean — and the ONLY package in the COMPOSITION tier. Its four packages mirror Gkeyll's model-family folders:

• mom/ owns five_moment, ten_moment, mhd, plasma (plasma parameters), multispecies (energetics, accumulate_current), rotations (par/perp to B), and the five-moment enstrophy/ke_dke frame programs.
• vm/ owns distribution-function frame transforms (kinetic) and particle trajectories (trajectory).
• pkpm/ owns Laguerre reconstruction and load_pkpm.
• gk/ owns distf/quantity loaders, the quantity registry (Tpar, beta, drift velocities), and energy_balance/particle_balance/nodes.

R-Z mapping and theta-phi flux-surface extraction are gyrokinetic operations; their old diagnostic module paths are compatibility aliases for the current major version. Contract: a diagnostic takes loaded data — one or several GData — plus physical scalars as keyword-only options, and returns GDataState (via _result, same inplace/tag/label contract as a verb) or a Figure; it is built entirely from the public vocabulary below it (operations, gdatastate, numerics, gdata) and nothing below the surfaces imports it. The render edge is pre-authorized for this layer (a program diagnostic may want render.plot()'s generic panel layout), but as of this writing every program module builds its own bespoke figure directly with matplotlib instead.

Each equation model owns its loading internally — there is no loaders/ package. Entry points like gk.load_quantity(...) (naming- convention load + registry dispatch, "physics-ready data by name") and pkpm.load_pkpm(...) live beside the physics they feed, because a quantity's ingredient files and its formula are one piece of equation knowledge. The only shared piece is diagnostics/discovery.py — equation-blind output-stem/frame discovery, the one home for Gkeyll's file-naming convention; equation loaders and programs resolve files through it, never with private globbing.

Functions have real names (mom.five_moment.pressure(d, gas_gamma=…)), never string dispatch; each equation module's VARIABLES table maps the CLI's quantity-name vocabulary ("density", "pressure", …) to those functions — the one home for that vocabulary. These are free functions, not GData methods: the layer sits above the fluent surface. (This layer absorbed the former models/ package — array physics now lives as private helpers inside the equation module that uses it.)

Engine layers — dg/, io/ (may import gpython only)

• dg/ — Gkeyll-kernel orchestration. dg/interpolate.py is the one-way modal→NumPy bridge (matrix from gpython.basis, applied per cell with tensordot; nodal-basis files convert through the exact nodal_to_modal matrix first); dg/modal.py holds the operations that stay modal (weak algebra, lincomb, shift_mean, power, integrate); dg/rep.py holds the explicit value_form changes (modal·nodal·quad) and apply_pointwise — all on native arrays.
• io/ — file I/O: read() dispatches over a reader registry. GkylCReader (first) reads field files entirely inside Gkeyll (gkyl_grid_array_new_from_file) and returns a native GkylArray; the pure-Python GkylReader is the fallback for no-library installs, partial loads, and dynvectors. save() supports gkyl/txt/npy/vtk. Readers fill a plain ctx dict and return (grid, values) — they never import gdatastate.

Leaves — numerics/ (imports nothing), gpython/ (the foreign floor)

• numerics/ — pure NumPy: idx_parser (selection strings) and elementwise (grids_compatible). No GData, ever.
• gpython/ — the only doorway to the foreign world (a test enforces this), and it is a compiled one (GKEYLL_C_SHIM.md): the gpython shim (gkeyll/core/zero/{gkyl_gpython.h, gpython.c}) lives in the gkeyll tree and is compiled by Gkeyll's own make core into libg0core.so — it holds every struct access, the by-value struct gkyl_basis convention, and the basis function-pointer dispatch, all checked by the C compiler against the headers in the same tree (shim and library can never drift apart). csrc/_gpythonmodule.c wraps gkyl_gpython.h (opaque handles + scalars + buffers only) into the _gpython extension, built by scripts/build_gpython.sh against the pinned gkeyll/ clone's libg0core.so, copies that library beside the exte </INSTRUCTIONS>

<environment_context> <cwd>/home/maxwell-rosen/postgkyl</cwd> <shell>zsh</shell> <current_date>2026-09-04</current_date> <timezone>America/New_York</timezone> <filesystem><workspace_roots><root>/home/maxwell-rosen/postgkyl</root></workspace_roots><permission_profile type="managed"><file_system type="restricted"><entry access="read"><special>:root</special></entry><entry access="write"><path>/home/maxwell-rosen/postgkyl</path></entry><entry access="write"><special>:slash_tmp</special></entry><entry access="write"><special>:tmpdir</special></entry><entry access="read"><path>/home/maxwell-rosen/postgkyl/.git</path></entry><entry access="read"><path>/home/maxwell-rosen/postgkyl/.agents</path></entry><entry access="read"><path>/home/maxwell-rosen/postgkyl/.codex</path></entry></file_system></permission_profile></filesystem> </environment_context>

Ruff is now installed. Run the relevant tests. Is there anything else which must be installed?

I installed these packages. Run the relevant tests

I got a big error running pip install -e '.[test]'

Requirement already satisfied: mdurl~=0.1 in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from markdown-it-py>=2.2.0->rich>=13.6.0->cyclopts>=4.0.0->pyvista>=0.48.4->postgkyl==2.0.0) (0.1.2) Requirement already satisfied: numexpr>=2.6.2 in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from tables>=3.10.1->postgkyl==2.0.0) (2.14.1) Requirement already satisfied: py-cpuinfo in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from tables>=3.10.1->postgkyl==2.0.0) (9.0.0) Requirement already satisfied: blosc2>=2.3.0 in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from tables>=3.10.1->postgkyl==2.0.0) (4.2.0) Requirement already satisfied: ndindex in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from blosc2>=2.3.0->tables>=3.10.1->postgkyl==2.0.0) (1.10.1) Requirement already satisfied: pydantic in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from blosc2>=2.3.0->tables>=3.10.1->postgkyl==2.0.0) (2.13.4) Requirement already satisfied: requests in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from blosc2>=2.3.0->tables>=3.10.1->postgkyl==2.0.0) (2.32.5) Requirement already satisfied: threadpoolctl in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from blosc2>=2.3.0->tables>=3.10.1->postgkyl==2.0.0) (3.6.0) Requirement already satisfied: readme-renderer>=35.0 in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from twine>=6.2.0->postgkyl==2.0.0) (46.0) Requirement already satisfied: requests-toolbelt!=0.9.0,>=0.8.0 in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from twine>=6.2.0->postgkyl==2.0.0) (1.0.0) Requirement already satisfied: urllib3>=1.26.0 in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from twine>=6.2.0->postgkyl==2.0.0) (2.6.3) Requirement already satisfied: keyring>=21.2.0 in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from twine>=6.2.0->postgkyl==2.0.0) (25.7.0) Requirement already satisfied: rfc3986>=1.4.0 in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from twine>=6.2.0->postgkyl==2.0.0) (2.0.0) Requirement already satisfied: id in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from twine>=6.2.0->postgkyl==2.0.0) (1.6.1) Requirement already satisfied: SecretStorage>=3.2 in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from keyring>=21.2.0->twine>=6.2.0->postgkyl==2.0.0) (3.5.0) Requirement already satisfied: jeepney>=0.4.2 in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from keyring>=21.2.0->twine>=6.2.0->postgkyl==2.0.0) (0.9.0) Requirement already satisfied: jaraco.classes in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from keyring>=21.2.0->twine>=6.2.0->postgkyl==2.0.0) (3.4.0) Requirement already satisfied: jaraco.functools in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from keyring>=21.2.0->twine>=6.2.0->postgkyl==2.0.0) (4.6.0) Requirement already satisfied: jaraco.context in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from keyring>=21.2.0->twine>=6.2.0->postgkyl==2.0.0) (6.1.2) Requirement already satisfied: nh3>=0.2.14 in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from readme-renderer>=35.0->twine>=6.2.0->postgkyl==2.0.0) (0.3.7) Requirement already satisfied: charset_normalizer<4,>=2 in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from requests->blosc2>=2.3.0->tables>=3.10.1->postgkyl==2.0.0) (3.4.6) Requirement already satisfied: idna<4,>=2.5 in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from requests->blosc2>=2.3.0->tables>=3.10.1->postgkyl==2.0.0) (3.11) Requirement already satisfied: certifi>=2017.4.17 in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from requests->blosc2>=2.3.0->tables>=3.10.1->postgkyl==2.0.0) (2026.2.25) Requirement already satisfied: cryptography>=2.0 in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from SecretStorage>=3.2->keyring>=21.2.0->twine>=6.2.0->postgkyl==2.0.0)(50.0.1) Requirement already satisfied: cffi>=2.0.0 in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from cryptography>=2.0->SecretStorage>=3.2->keyring>=21.2.0->twine>=6.2.0->postgkyl==2.0.0) (2.0.0) Requirement already satisfied: pycparser in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from cffi>=2.0.0->cryptography>=2.0->SecretStorage>=3.2->keyring>=21.2.0->twine>=6.2.0->postgkyl==2.0.0) (2.22) Requirement already satisfied: distlib<1,>=0.3.7 in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from virtualenv>=20.10.0->pre-commit==4.3.0->postgkyl==2.0.0) (0.4.3) Requirement already satisfied: filelock<4,>=3.24.2 in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from virtualenv>=20.10.0->pre-commit==4.3.0->postgkyl==2.0.0) (3.32.5) Requirement already satisfied: python-discovery>=1.6 in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from virtualenv>=20.10.0->pre-commit==4.3.0->postgkyl==2.0.0) (1.6.0) Requirement already satisfied: more-itertools in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from jaraco.classes->keyring>=21.2.0->twine>=6.2.0->postgkyl==2.0.0) (11.1.0) Requirement already satisfied: annotated-types>=0.6.0 in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from pydantic->blosc2>=2.3.0->tables>=3.10.1->postgkyl==2.0.0) (0.7.0) Requirement already satisfied: pydantic-core==2.46.4 in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from pydantic->blosc2>=2.3.0->tables>=3.10.1->postgkyl==2.0.0) (2.46.4) Requirement already satisfied: typing-inspection>=0.4.2 in /home/maxwell-rosen/miniforge3/lib/python3.13/site-packages (from pydantic->blosc2>=2.3.0->tables>=3.10.1->postgkyl==2.0.0) (0.4.2) Downloading pre_commit-4.3.0-py2.py3-none-any.whl (220 kB) Building wheels for collected packages: postgkyl Building editable for postgkyl (pyproject.toml) ... error error: subprocess-exited-with-error

× Building editable for postgkyl (pyproject.toml) did not run successfully. │ exit code: 1 ╰─> [326 lines of output] /tmp/pip-build-env-0sv6btj5/overlay/lib/python3.13/site-packages/setuptools/config/_apply_pyprojecttoml.py:82: SetuptoolsDeprecationWarning: project.license as a TOML table is deprecated !!

note: This error originates from a subprocess, and is likely not a problem with pip. ERROR: Failed building editable for postgkyl Failed to build postgkyl error: failed-wheel-build-for-install

× Failed to build installable wheels for some pyproject.toml based projects ╰─> postgkyl

How can we fix this bug?